Seismic processing prediction with a generative adversarial network

Abstract

Seismic processing involves a large number of analytic tools, that step by step, transform seismic field data into interpretable images of the subsurface. To obtain a seismic image that represents the geological model accurately, considerable time and computational resources are needed, as well as accurate physical models that are close to the actual solution. Unlike analytical methods, deep learning doesn’t need to have an exact knowledge of the physical behavior, changing how to make adjustments to the seismic signal. On the other hand, having a good amount of data set suitable for training the model becomes a limitation because the field data are limited and all are used for building a seismic image. In this work, we used Generative Adversarial Networks (GAN), initially designed to generate new images from a reference image, these types of networks don’t require a large amount of training data set and use low computational resources. In this work, two scenarios are presented. The first, proposes a new interpolation methodology applicable to OBN acquisitions, for 2D and 3D. This methodology describes the selection and preparation of training data and a workflow to train a GAN model and make a prediction using the same seismic data acquisition. The second scenario makes predictions of seismic images migrated with full-track processing, taking fast-track data as input to the model. The training methodology is performed conventionally, that is, a part of the data was used for training and another for testing, obtaining an efficient model in its predictions with the possibility of being used with time-lapse data. For both proposed scenarios, the same GAN is applied, modifying the training data for each case, demonstrating the flexibility of this type of network performing different tasks related to seismic processing with a small amount of data.